[go: up one dir, main page]
More Web Proxy on the site http://driver.im/

CN113066544B - FVEP characteristic point detection method based on CAA-Net and LightGBM - Google Patents

FVEP characteristic point detection method based on CAA-Net and LightGBM Download PDF

Info

Publication number
CN113066544B
CN113066544B CN202110434370.6A CN202110434370A CN113066544B CN 113066544 B CN113066544 B CN 113066544B CN 202110434370 A CN202110434370 A CN 202110434370A CN 113066544 B CN113066544 B CN 113066544B
Authority
CN
China
Prior art keywords
sequence
points
fvep
feature point
feature
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202110434370.6A
Other languages
Chinese (zh)
Other versions
CN113066544A (en
Inventor
陈娟娟
谢辛
汪成亮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing Normal University
Original Assignee
Chongqing Normal University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing Normal University filed Critical Chongqing Normal University
Priority to CN202110434370.6A priority Critical patent/CN113066544B/en
Publication of CN113066544A publication Critical patent/CN113066544A/en
Application granted granted Critical
Publication of CN113066544B publication Critical patent/CN113066544B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
    • G16H10/20—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for electronic clinical trials or questionnaires
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
    • G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Epidemiology (AREA)
  • Primary Health Care (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a method for detecting FVEP characteristic points based on CAA-Net and LightGBM, which comprises the following steps: s1, preprocessing data: standardizing the FVEP signal as an input of a neural network model; s2, generating a characteristic point sequence: s2-1, selecting position coordinates which are possibly characteristic points through a CAA-Net model, and filtering the characteristic points to be selected which are less likely; s2-2, generating a feature point sequence to be selected: if the points in the feature point set to be selected are not extreme points, searching in the range that the distance of the feature points in the feature point set to be selected is less than 2, if the extreme points cannot be found, discarding the points, and if the extreme points are found, replacing the feature points to be selected with the extreme points; then, generating a preliminary feature point sequence set to be selected according to an exhaustion method; and S3, selecting the characteristic point sequence to obtain the optimal characteristic point sequence. The invention can quickly and effectively obtain the FVEP waveform characteristic points, thereby carrying out clinical analysis on the illness state of the patient.

Description

FVEP characteristic point detection method based on CAA-Net and LightGBM
Technical Field
The invention relates to the field of medical data detection, in particular to a method for detecting FVEP characteristic points based on CAA-Net and LightGBM.
Background
The FVEP waveform characteristic point is an important basis for the clinician to judge the state of an illness; however, the clinical analysis and interpretation are difficult due to poor patient matching, large individual difference of the FVEP waveform, large intra-individual variation and the like. Therefore, research on the FVEP waveform characteristic point is an important subject to be solved urgently.
Disclosure of Invention
The invention aims to at least solve the technical problems in the prior art, and particularly creatively provides a method for detecting FVEP characteristic points based on CAA-Net and LightGBM.
In order to achieve the above object, the present invention provides a method for detecting FVEP characteristic points based on CAA-Net and LightGBM, comprising the following steps:
s1, preprocessing data: standardizing the FVEP signal as an input of a neural network model;
s2, generating a characteristic point sequence:
s2-1, selecting position coordinates which are possibly characteristic points through a CAA-Net model, and filtering the characteristic points to be selected which are less likely;
s2-2, generating a feature point sequence to be selected: if the points in the feature point set to be selected are not extreme points, searching in the range that the distance of the feature points in the feature point set to be selected is less than 2, if the extreme points cannot be found, discarding the points, and if the extreme points are found, replacing the feature points to be selected with the extreme points; then, generating a preliminary feature point sequence set to be selected according to an exhaustion method;
and S3, selecting the characteristic point sequence to obtain the optimal characteristic point sequence.
In a preferred embodiment of the present invention, the S1 includes:
z-score normalization of the FVEP data: for each piece of FVEP signal data, the data in 320 milliseconds of the examined person is included, wherein one piece of data is sampled every 1 millisecond and forms a sequence
Figure BDA0003032343620000021
Wherein,
Figure BDA0003032343620000022
data representing the 1 st millisecond of sampling,
Figure BDA0003032343620000023
data representing the k-1 millisecond sample，
Figure BDA0003032343620000024
Data representing the k-th millisecond sample is,
Figure BDA0003032343620000025
data representing the k +1 millisecond sample,
Figure BDA0003032343620000026
data representing a 320 millisecond sample; calculating XiMean value of
Figure BDA0003032343620000027
XiThe standard deviation s of (a) is transformed as follows:
Figure BDA0003032343620000028
wherein,
Figure BDA0003032343620000029
representing the value of the ith FVEP waveform sample normalized by the kth millisecond.
In a preferred embodiment of the present invention, the S2-1 includes:
the CAA-Net firstly extracts local features from the input FVEP signals through one-dimensional convolution, and then reserves the most useful features through the maximum pooling; adding a one-dimensional convolution maximum pooling layer; the feature diagram output at this time is sent to a self-attention layer; outputting the self-attention layer, and performing one-dimensional convolution to obtain a characteristic diagram; adding two branches, wherein one branch is added with a one-dimensional convolution and combined with a sigmoid activation function to extract the probability of the characteristic point, and the other branch is added with a one-dimensional convolution and combined with a linear activation function to extract the offset of the position;
the calculation formula of the bias is as follows:
Figure BDA00030323436200000210
Figure BDA00030323436200000211
wherein x is the coordinate of the input point, xoutFor coordinates through the feature map, xbias is the bias [ ·]Meaning rounding down, |, denotes the absolute value.
In a preferred embodiment of the present invention, the biasing comprises:
to facilitate neural network processing, the bias is normalized, namely xbias/4.
In a preferred embodiment of the present invention, the S2-1 includes:
a one-dimensional convolution and self-attention mechanism is used: for extracting local features in the waveform, one-dimensional convolution is adopted; for the relation between the search characteristic point and other characteristic points, an attention mechanism is adopted.
In a preferred embodiment of the present invention, the S3 includes the steps of:
s3-1, calculating probability of the characteristic point sequences to be selected by using a multivariate Gaussian model, expressing the position probability of the characteristic point sequences to be selected, filtering low-probability sequences to be selected, and reserving M sequences to be selected with the highest probability;
s3-2, calculating the form probability and the feature point probability of the sequence to be selected, combining the position probability, the age and the m feature point coordinates to generate feature point sequence features, predicting by adopting a LightGBM model, and predicting the average absolute error between the sequence to be selected and the real sequence.
And S3-3, selecting the candidate sequence with the lowest predicted average absolute error from the M candidate sequences as the optimal characteristic point sequence.
In a preferred embodiment of the present invention, the multivariate gaussian model comprises:
Figure BDA0003032343620000031
wherein y represents a sequence of feature pointsMu represents the mean vector, sigma represents the covariance matrix, | - | represents the value of the determinant in linear algebra, ·TRepresents a transposition, ·-1Representing the inverse matrix and d representing the characteristic point sequence dimension. The invention leads d to be 6. And carrying out maximum likelihood estimation by utilizing the training set data, and calculating multivariate Gaussian distribution parameters mu and sigma. When a new sequence of feature points is entered, its probability density can be calculated.
In a preferred embodiment of the present invention, the selecting of the feature point sequence in S3 includes:
in order to detect the feature point sequence, the selected optimal feature point sequence is closest to the real feature point sequence, and the average absolute error MAE between the sequence to be selected and the real feature point sequence is made as follows:
Figure BDA0003032343620000032
wherein Y iskRepresents the k-th true feature point,
Figure BDA0003032343620000033
and (4) representing the kth candidate characteristic point, |, representing an absolute value.
In a preferred embodiment of the invention, the mean absolute error of the sequence is:
Figure BDA0003032343620000041
sequence root mean square error:
Figure BDA0003032343620000042
mean absolute percent error of sequence:
Figure BDA0003032343620000043
wherein,
Figure BDA0003032343620000044
represents the predicted amount of the ith FVEP signal sampling time k,
Figure BDA0003032343620000045
and the mark representing that the sampling time of the ith FVEP signal is k, n is the number of samples of the data set, m is the vector dimension of the characteristic point, namely the number of the characteristic points, | · | represents an absolute value.
In summary, due to the adoption of the technical scheme, the invention has the beneficial effects that: the method can quickly and effectively obtain the FVEP waveform characteristic points, thereby carrying out clinical analysis on the illness state of the patient.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Drawings
The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 is a schematic diagram of ADFP model training of the present invention.
FIG. 2 is a schematic diagram of ADFP model prediction in accordance with the present invention.
FIG. 3 is a schematic diagram of the CAA-Net structure of the present invention.
FIG. 4 is a schematic diagram of the probability of a calculated sequence of the multivariate Gaussian model of the present invention.
FIG. 5 is a schematic diagram of the probability visualization of the FVEP morphology according to the present invention.
FIG. 6 is a schematic representation of the features and labels of the present invention.
Fig. 7 is a feature point histogram of the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like or similar reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative only for the purpose of explaining the present invention, and are not to be construed as limiting the present invention.
The characteristic points of the FVEP waveform are an important basis for the clinician to judge the condition of the patient. The invention provides a framework ADFP for automatically detecting FVEP signal characteristic points, which is cascaded with a neural network model CAA-Net, a multivariate Gaussian model and a LightGBM model based on a convolutional neural network and an attention mechanism.
The ADFP model provided by the invention can be divided into the following three steps in the detection process of the FVEP characteristic points:
and S1, preprocessing. Normalization of the FVEP signal is used in this step as an input to the neural network model. Preprocessing section as in fig. 2
And S2, generating a characteristic point sequence. In the step, position coordinates which are possibly characteristic points are selected through a CAA-Net model, and the characteristic points to be selected with low possibility are filtered. If the points in the feature point set to be selected are not extreme points, searching in the range that the distance of the feature points in the feature point set to be selected is less than 2, if the extreme points cannot be found, discarding the points, and if the extreme points are found, replacing the feature points to be selected with the extreme points. And then, generating a preliminary feature point sequence set to be selected according to an exhaustion method. Such as the feature point sequence generation portion of fig. 2.
S3, feature point sequence selection. Firstly, calculating probability of a feature point sequence to be selected by using a multivariate Gaussian model, expressing the position probability of the feature point sequence, filtering low-probability sequences to be selected, and reserving M sequences to be selected with highest probability. And then, calculating the form probability and the feature point probability of the sequence to be selected, combining the position probability, the age and the m feature point coordinates to generate feature point sequence features, predicting by adopting a LightGBM model, and predicting the average absolute error between the sequence to be selected and the real sequence. And finally, selecting the candidate sequence with the lowest predicted average absolute error from the M candidate sequences as the optimal characteristic point sequence. As shown in FIG. 2, the feature point sequence selecting section
The training process of the ADFP model is shown in fig. 1, and the prediction process is shown in fig. 2, wherein the specific implementation of each given step is as follows:
1. problem definition
For the input training set R { (X)1,Y1),...,(Xi,Yi),...,(XD,YD)}. D is the number of samples in the training set, XiRepresents a FVEP signal, YiRepresents m characteristic points
Figure BDA0003032343620000061
Will be trained in the training set R to obtain the model ADFP, so that the new FVEP signal X is inputnewOutput Ynew。
2. Data pre-processing
In clinic, the magnitude of the FVEP signals of different examinees is different greatly, but the FVEP signals are of little significance to the doctor to judge the disease condition. To eliminate the interference of such magnitudes with the neural network model, the FVEP data was z-score normalized. For each piece of FVEP signal data, the data in 320 milliseconds of the examined person is included, wherein one piece of data is sampled every 1 millisecond and forms a sequence
Figure BDA0003032343620000062
Wherein,
Figure BDA0003032343620000063
data representing the 1 st millisecond of sampling,
Figure BDA0003032343620000064
data representing the k-1 millisecond sample,
Figure BDA0003032343620000065
data representing the k-th millisecond sample is,
Figure BDA0003032343620000066
data representing the k +1 millisecond sample,
Figure BDA0003032343620000067
data representing a 320 millisecond sample; calculating XiMean value of
Figure BDA0003032343620000068
XiThe standard deviation s of (a) is transformed as follows:
Figure BDA0003032343620000069
wherein,
Figure BDA00030323436200000610
representing the value of the ith FVEP waveform sample normalized by the kth millisecond.
CAA-Net network
The method for automatically detecting the m characteristic points in the FVEP signal is greatly different from the traditional target detection task. In the traditional target detection task, each detected target is independent of other targets, and the sequence among the detected targets does not need to be considered. In the detection task of the FVEP feature point, the sequence is predicted
Figure BDA00030323436200000611
When m is taken as 6, the compound is,
Figure BDA00030323436200000612
representing six characteristic points as a sequence, wherein each point is required to depend on the existence of other points, and the characteristic points N are all troughs, including the characteristic points N1, N2 and N3; the characteristic points P are all wave crests and comprise characteristic points P1, P2 and P3. This feature distinguishes the FVEP feature detection task from the general object detection task.
Based on a sequence-to-sequence neural network model CAA-Net, an input is an FVEP signal, the probability that characteristic points possibly exist at each position is output, and the structure of the proposed CAA-Net model is shown in FIG. 3. The CAA-Net firstly extracts local features from the input FVEP signal through a layer of one-dimensional convolution Con1D, and then retains the most useful features through a maximum pooling layer MaxPool. In order to obtain a deeper receptive field, a one-dimensional convolution and maximum pooling layer Con1D-MaxPool structure is added. The output characteristic diagram is sent to a self-attention layer. Using the self-attention layer, each vector in the feature map will be explored for relationships with other vectors. And obtaining a feature map through one-dimensional convolution Con1D on the output of the attention layer. Consider the use of two times maximum pooling throughout the neural network structure, the output sequence length dimension of the feature map 80. The output vector corresponding to each feature map corresponds to 4 points of the original input. At the moment, two branches are added, wherein one branch is added with one-dimensional convolution and combined with a sigmoid activation function to extract the feature point probability, and the other branch is added with one-dimensional convolution and combined with a linear activation function to extract the offset of the position.
Considering that the input sequence length is 320, the output feature point probability sequence length is 80, that is, the probability of each output probability sequence represents the probability that a feature point exists in the range of the input sequence with the receptive field of 4, so that the accurate corresponding position of the input sequence cannot be obtained. To solve this problem, a bias branch is added. Assuming that the coordinate of the input point is x, obtaining the coordinate x through the characteristic diagramoutThe bias is calculated as xbias, as shown in equation 3.3, as shown in equation 3.4. To facilitate neural network processing, the bias is normalized, namely xbias/4.
Figure BDA0003032343620000071
Figure BDA0003032343620000072
Wherein x is the coordinate of the input point, xoutFor coordinates through the feature map, xbias is the bias [ ·]Expressing rounding down, | represents an absolute value;
for selecting the feature points, two factors are considered, namely whether the feature points are extreme points and morphological features thereof, and the positions of the feature points in the whole sequence and the relationship with other feature points. To solve this problem, a one-dimensional convolution and a self-attention mechanism are used. For extracting local features in the waveform, one-dimensional convolution is adopted. Similar to images, where the former convolutional layer processes low-level features and the latter convolutional layer processes high-level features, only some local features with a small lower receptive field range need to be extracted in our task. For the relation between the searched feature point and other feature points, an attention mechanism is adopted, each vector in the feature map is subjected to inner product with all vectors in the feature map, and the global relation is searched.
CAA-Net has two functions in the ADFP model. The first function is the feature point probability output by the CAA-Net model, the feature point probability is filtered firstly, and when the feature point probability is larger than a threshold value tau, the feature point probability is added into a feature point set to be selected. If the feature points in the feature point set to be selected are not extreme points, searching in the range that the distance of the feature points in the feature point set to be selected is less than 2, if the extreme points cannot be found, discarding the points, and if the extreme points are found, replacing the feature points with the extreme points. And then traversing the set of the feature points to be selected to generate a set of a sequence of the feature points to be selected. The second function is to define the average value of the CAA-Net output probability in the feature point sequence to be selected as the feature point probability pcAnd taking the selected feature point as a feature of the feature point sequence to be selected.
4. Multivariate Gaussian model
The invention uses a multivariate Gaussian distribution model to model the characteristic point sequence. First, a multivariate gaussian distribution P (y, μ, Σ) is defined, where y represents a feature point sequence, μ represents a mean vector, and Σ represents a covariance matrix, as shown in equation 3.5:
Figure BDA0003032343620000081
where | represents the value of a determinant in linear algebra, ·TRepresents a transposition, ·-1Denotes an inverse matrix, d denotes a feature point sequence dimension, and in the present invention, d is 6. And carrying out maximum likelihood estimation by utilizing the training set data, and calculating multivariate Gaussian distribution parameters mu and sigma. When a new sequence of feature points is entered, its probability density can be calculated. As shown in fig. 4, after the multivariate gaussian distribution model is established, the probability densities of the three feature point sequences are calculated. Wherein the first sequence is the marker sequence whose probability density is the greatest. The second sequence is positioned closer to the marker sequence than the third sequence. It can be observed that the second sequence is almostThe rate density is also much higher than the probability density of the third sequence. Therefore, we define the probability density of a sequence as the sequence position probability pα。
The multivariate gaussian distribution model has two roles in the ADFP model. The first function is to carry out preliminary screening on a characteristic point sequence set generated by the CAA-Net model and reserve the first M sequences to be selected with the highest position probability. The second role is that it is a feature of the sequence of feature points as input to the LightGBM model.
5. Form probability
In the ADFP model, a plurality of factors are simultaneously considered for feature point detection, and the shapes of peaks and troughs are only taken as one of the considered features. The invention provides an algorithm for calculating the form probability of a peak and a trough to be selected, which is to calculate all possible peaks and troughs, namely all extreme points, of an FVEP signal. Then, for each extreme point, calculating the average value of the absolute value range in the range of 2 neighborhoods left and right of the extreme point, and then normalizing the average value of the range by using a maximum and minimum normalization algorithm.
As shown in fig. 5, the morphological probability of the FVEP signal extreme point is calculated by an algorithm. It is observed that the morphological probability of the feature point is not necessarily the highest, for example, the morphological probability of the N1 point is very small, and is only 0.04. Therefore, the morphology probability can only be used as one of the criteria for determining whether the feature point is present. And for a feature point sequence to be selected, defining the average morphological feature point probability as the feature point sequence morphological probability.
In the ADFP model, the morphological probability of the feature point sequence to be selected is used as one of the bases for judging whether the feature point sequence is a feature point. When the acquired FVEP data has high quality, the morphology probability plays a very important role in the acquired FVEP data, because the morphology difference between the characteristic point and the non-characteristic point is very obvious in the extreme point. In contrast, when the quality of the acquired FVEP data is not high, the morphological probability becomes relatively low therein.
6. Optimal sequence selection
Abstracting the optimal sequence selection problem, and assuming that N FVEP feature point sequences exist in a training set. And (3) processing by using a CAA-Net model and a multivariate Gaussian model, wherein each FVEP characteristic point sequence corresponds to M sequences to be selected, and the optimal sequence selection corresponds to the optimal selection from the M sequences. Our goal is to detect the sequence of feature points such that the selected optimal sequence is closest to the true sequence of feature points. Defining the average absolute error MAE between the sequence to be selected and the real characteristic point sequence:
Figure BDA0003032343620000101
wherein Y iskRepresents the k-th true feature point,
Figure BDA0003032343620000102
represents the kth candidate feature point, | · | represents an absolute value, and Y ═ Y1,Y2,Y3,Y4,Y5,Y6) Representing a sequence of true feature points, Y1Represents the 1 st true feature point, Y2Represents the 2 nd true feature point, Y3Represents the 3 rd true feature point, Y4Represents the 4 th true feature point, Y5Represents the 5 th true feature point, Y6Representing the 6 th real feature point;
Figure BDA0003032343620000103
the sequence to be selected is represented by a sequence,
Figure BDA0003032343620000104
the 1 st candidate feature point is represented,
Figure BDA0003032343620000105
the 2 nd candidate feature point is represented,
Figure BDA0003032343620000106
the 3 rd candidate feature point is represented,
Figure BDA0003032343620000107
the 4 th candidate feature point is represented,
Figure BDA0003032343620000108
the 5 th candidate feature point is represented,
Figure BDA0003032343620000109
and representing the 6 th candidate characteristic point. How to predict the MAE between the sequence to be selected and the real characteristic point sequence can be converted into a regression problem, namely, a regression model is established, the input is the sequence to be selected, the position probability, the form probability and the characteristic point probability, and the output is the MAE between the sequence to be selected and the real characteristic point sequence. The mean absolute value error is chosen as the predicted value rather than the mean square error MSE, because MSE is more sensitive to outliers. If one of the six feature points has a large error, the MSE of the whole sequence is large. We need a more stable, less sensitive loss evaluation method to outliers, and therefore choose MAE. Finally, the age, the form probability, the position probability, the feature point probability and the six feature points are selected, and a model for predicting the MAE is established. As shown in fig. 6.
After the characteristic variables and the predictive variables are determined, a machine learning regression model is needed. Common regression models include linear regression, KNN regression, support vector regression, and fusion-based regression models. LightGBM was chosen as the regression model, which is a high performance framework for implementing the GBDT algorithm. GBDT belongs to a method in ensemble learning, and the core idea is to adopt weak classifiers (such as decision trees) to iterate step by step so as to obtain an optimal model, and the method has the advantages of stable training performance, strong generalization capability, difficulty in overfitting and the like. The XGBOOST framework, also belonging to the GBDT framework, was proposed for use before LightGBM, which utilizes a pre-ordering technique based on a decision tree algorithm. The advantage is that the segmentation point of the decision tree is found accurately, but its disadvantages are also apparent: it is spatially complex because it requires a large number of intermediate results to be saved during the training process. In addition, because XGBOOST requires computation of the splitting gain each time it traverses, its time complexity is also high. To solve the XGBOOST problem, LightGBM employs some improved techniques including unilateral gradient sampling, GOSS, mutually exclusive feature bundling, EFB, and Leaf growth strategy with depth-limited Leaf-wise while ensuring accuracy.
In the ADFP model, LightGBM functions to select an optimal feature point sequence from M candidate feature point sequences. Firstly, predicting the MAE value of the characteristic point sequence to be selected by utilizing the LightGBM, and then selecting the sequence to be selected with the minimum predicted MAE value as the optimal sequence.
In the ADFP, the whole steps are divided into preprocessing, feature point sequence generation and feature point sequence selection. Firstly, generating feature points to be selected through a CAA-Net model. And then generating a set of to-be-selected feature point sequences, and selecting M to-be-selected sequences by using a multivariate Gaussian distribution model. Finally, the optimal sequence is selected using the LightGBM model.
7. Data set
The FVEP normal data set comprises 5164 pieces of FVEP data, and each person provides two detection results for the left and right eyes from 1366 examined persons. All FVEP data were collected in total at the first hospital affiliated of the army medical university (southwest hospital) ophthalmology over a time span from 7/1/2012 to 3/1/2020 using Espion E2 eye electrophysiological equipment loaded with the Espion E2 system.
Wherein, according to the ID grouping of the examinees, the data of 70% of the examinees are randomly selected as training set data, and the data of 30% of the examinees are selected as test set data. The data of each examined person includes the FVEP waveform data, the position of the feature point, the sex, the age, and the disease type. In the experiment of the present invention, three types of data, waveform data, feature point positions, and ages, will be used.
8. Evaluation index
The final output of the model of the invention is a sequence which comprises six characteristic points. By comparing and evaluating the output sequence with the real sequence, the capability of the model for detecting the characteristic points can be visually seen. Hypothesis prediction sequence
Figure BDA0003032343620000111
Wherein
Figure BDA0003032343620000112
Representing the ith sample with a feature point vector dimension of 1Measuring;
Figure BDA0003032343620000113
representing the prediction quantity of the ith sample with the feature point vector dimension of 2;
Figure BDA0003032343620000114
representing the prediction quantity of the ith sample with the feature point vector dimension of m-1;
Figure BDA0003032343620000115
and representing the prediction quantity with the dimension of m of the feature point vector of the ith sample. Marker sequences
Figure BDA0003032343620000116
Wherein,
Figure BDA0003032343620000117
the 1 st marker representing the ith sample,
Figure BDA0003032343620000118
the 2 nd marker representing the ith sample,
Figure BDA0003032343620000121
the mth label representing the ith sample,
Figure BDA0003032343620000122
an mth label representing the ith sample; in order to realize the performance evaluation of the established model, the model performance is evaluated by using the following four indexes:
mean absolute error of sequence SMAE: represents the average of the absolute errors between the predicted sequence and the marker sequence. The formula is as follows:
Figure BDA0003032343620000123
sequence root mean square error SRMSE: representing the square root of the mean of the squared differences between the predicted sequence and the marker sequence. The formula is as follows:
Figure BDA0003032343620000124
mean absolute percent error of sequence SMAPE: representing the sum of the absolute errors between the predicted and actual observations divided by the actual value. In practice, it is the average of the error percentages.
Figure BDA0003032343620000125
Wherein, Yi kRepresents the predicted amount of the ith FVEP signal sampling time k,
Figure BDA0003032343620000126
and the mark representing that the sampling time of the ith FVEP signal is k, n is the number of samples of the data set, m is the vector dimension of the characteristic point, namely the number of the characteristic points, | · | represents an absolute value.
Average run time: indicating the average time of the FVEP from pre-processing to outputting the sequence of feature points in the data set.
9. Experimental parameters
All neural network models use the RAdam optimizer and the cosine annealing learning rate annealing algorithm. Wherein the learning rate of the RAdam optimizer is set to 0.01, the batch size is set to 128, and the number of iteration rounds is 50. The cosine learning rate annealing algorithm linearly increases the selection learning rate in the first 10 rounds and decreases the cosine learning rate in the 10 th to 50 th rounds. All neural network models have a Loss function of Focal local. And setting a threshold value tau to be 0.15 for the probability of the characteristic point output by the neural network, and filtering the candidate characteristic points with the output probability smaller than tau. For the neural network, the model with the lowest training loss in the training set is selected as the optimal model. For the multivariate gaussian model, the M value is set to 10. For the LightGBM model, the maximum leaf node number is set to be 30, the maximum depth is set to be 10, the learning rate is 0.02, the maximum cake number is set to be 50, and 80% of training sets are selected in each iteration.
10. Results and analysis of the experiments
TABLE 3.2 comparison of test model Performance
Figure BDA0003032343620000131
The experiment of the invention compares the ADFP model with other models, including a multivariate Gaussian model, an LSTM model and a CAA-Net model. As can be seen from table 3.2, the run time is very slow, thirty times that of the ADFP model, using only the multivariate gaussian model. Analysis shows that the multivariate Gaussian model needs to traverse all extreme points, the search space is huge, and a large amount of computing resources are consumed. Meanwhile, only the factors of the positions of the feature points are considered, and the features of the form, the age and the like are not considered, so that the prediction accuracy is low. With the wavelet transform, the run time is faster, and the search space is greatly reduced because it filters out some of the noise. In the combination of the multivariate Gaussian model and the two network structures, the CAA-Net network has the highest prediction accuracy, and because the self-attention mechanism is utilized to find the relation between the characteristic points and other characteristic points, the multivariate Gaussian model has an auxiliary effect on selecting better characteristic points.
Compared with other models, the ADFP model (without adding the age characteristics) has obviously improved prediction performance. Compared with the CAA-Net model, the SRMSE is reduced by 0.67, the SAME is reduced by 0.77, and the SMAPE is reduced by 0.18, which shows that the optimal sequence selection is efficient by utilizing various characteristics and combining LightGBM. When the ADFP model increases the age characteristic, the SRMSE of the ADFP model is reduced by 0.68, the SMAE is reduced by 0.46, the SMAPE is reduced by 0.46, the performance of the ADFP model is obviously improved, and the importance of the age characteristic in the characteristic point detection is also explained.
TABLE 3.4 Performance of different feature points of the ADFP model
Characteristic point MAE Median absolute error
N1 3.86 1.0
P1 6.70 1.0
N2 8.54 1.0
P2 7.64 1.0
N3 12.51 5.0
P3 19.48 12.0
Table 3.4 calculates the mean absolute error MAE for each feature point in the predicted sequence, as well as the median of the absolute errors. As can be seen from table 3.4, the ADFP model showed different performance for the prediction performance of each feature point. The predicted performance of four points of N1, P1, N2 and P2 is better, and the MAE of the four points is less than 10. Of the six feature points, except for the point N2 and the point P2, the closer to the back of the sequence, the poorer the prediction performance. Point N2 was more advanced in the sequence than point P2, but its MAE value was 0.90 higher than that at point P2. As shown in fig. 7, the distribution of feature points in the dataset is similar to gaussian distribution except for point N2, but point N2 has two peaks, and the distribution is more discrete with respect to point P2, which is the reason for its lower predictive performance. MAE is the mean of absolute errors, which we can find to be much higher than the median of absolute errors, indicating that there are some FVEP sequences with poor prediction performance. In clinical application, due to the variability of the FVEP, the characteristic points N3 and P3 are not used as diagnostic reference standards of doctors, and the focused performance index is mainly the characteristic point P2.
While embodiments of the invention have been shown and described, it will be understood by those of ordinary skill in the art that: various changes, modifications, substitutions and alterations can be made to the embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims (7)

1. A method for detecting FVEP feature points based on CAA-Net and LightGBM is characterized by comprising the following steps:
s1, preprocessing data: standardizing the FVEP signal as an input of a neural network model;
s2, generating a characteristic point sequence:
s2-1, selecting position coordinates which are possibly characteristic points through a CAA-Net model, and filtering the characteristic points to be selected which are less likely;
the CAA-Net firstly extracts local features from the input FVEP signals through one-dimensional convolution, and then reserves the most useful features through the maximum pooling; adding a one-dimensional convolution maximum pooling layer; the feature diagram output at this time is sent to a self-attention layer; outputting the self-attention layer, and performing one-dimensional convolution to obtain a characteristic diagram; adding two branches, wherein one branch is added with a one-dimensional convolution and combined with a sigmoid activation function to extract the probability of the characteristic point, and the other branch is added with a one-dimensional convolution and combined with a linear activation function to extract the offset of the position;
the calculation formula of the bias is as follows:
Figure FDA0003525520830000011
Figure FDA0003525520830000012
wherein x is the coordinate of the input point, xoutFor coordinates through the feature map, xbias is the bias [ ·]Expressing rounding down, | represents an absolute value;
s2-2, generating a feature point sequence to be selected: if the points in the feature point set to be selected are not extreme points, searching in the range that the distance of the feature points in the feature point set to be selected is less than 2, if the extreme points cannot be found, discarding the points, and if the extreme points are found, replacing the feature points to be selected with the extreme points; then, generating a preliminary feature point sequence set to be selected according to an exhaustion method;
s3, selecting the characteristic point sequence to obtain an optimal characteristic point sequence;
s3-1, calculating probability of the characteristic point sequences to be selected by using a multivariate Gaussian model, expressing the position probability of the characteristic point sequences to be selected, filtering low-probability sequences to be selected, and reserving M sequences to be selected with the highest probability;
s3-2, calculating the form probability and the feature point probability of the sequence to be selected, combining the position probability, the age and the m feature point coordinates to generate feature point sequence features, predicting by adopting a LightGBM model, and predicting the average absolute error between the sequence to be selected and the real sequence;
s3-3, selecting a candidate sequence with the lowest predicted average absolute error from the M candidate sequences as an optimal characteristic point sequence;
the CAA-Net is a sequence-to-sequence neural network model.
2. The method of claim 1, wherein the step S1 comprises:
z-score normalization of the FVEP data: for each piece of FVEP signal data, the data in 320 milliseconds of the examined person is included, wherein one piece of data is sampled every 1 millisecond and forms a sequence
Figure FDA0003525520830000021
Wherein,
Figure FDA0003525520830000022
data representing the 1 st millisecond of sampling,
Figure FDA0003525520830000023
data representing the k-1 millisecond sample,
Figure FDA0003525520830000024
data representing the k-th millisecond sample is,
Figure FDA0003525520830000025
data representing the k +1 millisecond sample,
Figure FDA0003525520830000026
data representing a 320 millisecond sample; calculating XiMean value of
Figure FDA0003525520830000027
XiThe standard deviation s of (a) is transformed as follows:
Figure FDA0003525520830000028
wherein,
Figure FDA0003525520830000029
representing the value of the ith FVEP waveform sample normalized by the kth millisecond.
3. The method of claim 1, wherein the biasing comprises:
the bias is normalized, i.e., xbias/4.
4. The method for detecting the FVEP feature point based on CAA-Net and LightGBM of claim 1, wherein the S2-1 comprises:
a one-dimensional convolution and self-attention mechanism is used: for extracting local features in the waveform, one-dimensional convolution is adopted; for the relation between the search characteristic point and other characteristic points, an attention mechanism is adopted.
5. The method of claim 1, wherein the multivariate Gaussian model comprises:
Figure FDA0003525520830000031
where y represents a sequence of feature points, μ represents a mean vector, Σ represents a covariance matrix, | · | represents the value of a determinant in a linear algebra, ·TRepresents a transposition, ·-1Representing the inverse matrix and d representing the characteristic point sequence dimension.
6. The method of claim 1, wherein the step of selecting the feature point sequence in S3 comprises:
mean absolute error MAE between the candidate sequence and the true feature point sequence:
Figure FDA0003525520830000032
wherein Y iskRepresents the k-th true feature point,
Figure FDA0003525520830000033
and (4) representing the kth candidate characteristic point, |, representing an absolute value.
7. The method for detecting the FVEP feature point based on CAA-Net and LightGBM as claimed in claim 1, further comprising an evaluation index:
mean absolute error of sequence:
Figure FDA0003525520830000034
sequence root mean square error:
Figure FDA0003525520830000035
mean absolute percent error of sequence:
Figure FDA0003525520830000041
wherein, Yi kRepresents the predicted amount of the ith FVEP signal sampling time k,
Figure FDA0003525520830000042
and the mark representing the sampling time k of the ith FVEP signal, n is the number of samples of the data set, m is the dimension of the feature point vector, | · | represents an absolute value.
CN202110434370.6A 2021-04-22 2021-04-22 FVEP characteristic point detection method based on CAA-Net and LightGBM Active CN113066544B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110434370.6A CN113066544B (en) 2021-04-22 2021-04-22 FVEP characteristic point detection method based on CAA-Net and LightGBM

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110434370.6A CN113066544B (en) 2021-04-22 2021-04-22 FVEP characteristic point detection method based on CAA-Net and LightGBM

Publications (2)

Publication Number Publication Date
CN113066544A CN113066544A (en) 2021-07-02
CN113066544B true CN113066544B (en) 2022-04-15

Family

ID=76567378

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110434370.6A Active CN113066544B (en) 2021-04-22 2021-04-22 FVEP characteristic point detection method based on CAA-Net and LightGBM

Country Status (1)

Country Link
CN (1) CN113066544B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115512717A (en) * 2022-11-21 2022-12-23 东北石油大学三亚海洋油气研究院 Two-stage single-sensor pipeline leakage positioning method based on ensemble learning

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108664967A (en) * 2018-04-17 2018-10-16 上海交通大学 A kind of multimedia page vision significance prediction technique and system
CN110929918A (en) * 2019-10-29 2020-03-27 国网重庆市电力公司南岸供电分公司 10kV feeder line fault prediction method based on CNN and LightGBM
CN111177970A (en) * 2019-12-10 2020-05-19 浙江大学 Multi-stage semiconductor process virtual metering method based on Gaussian process and convolutional neural network
CN112487291A (en) * 2020-11-28 2021-03-12 重庆邮电大学 Big data-based personalized news recommendation method and device

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108664967A (en) * 2018-04-17 2018-10-16 上海交通大学 A kind of multimedia page vision significance prediction technique and system
CN110929918A (en) * 2019-10-29 2020-03-27 国网重庆市电力公司南岸供电分公司 10kV feeder line fault prediction method based on CNN and LightGBM
CN111177970A (en) * 2019-12-10 2020-05-19 浙江大学 Multi-stage semiconductor process virtual metering method based on Gaussian process and convolutional neural network
CN112487291A (en) * 2020-11-28 2021-03-12 重庆邮电大学 Big data-based personalized news recommendation method and device

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
CAA-Net:Conditional Atrous CNNs With Attention for Explainable Device-Robust Acoustic Scene Classification;Zhao Ren 等;《IEEE Transactions on Multimedia》;20201116;第23卷;4131-4142 *
基于C-lightGBM的用户窃电检测;刘海青 等;《计算机应用研究》;20200630;第37卷(第S1期);298-300,303 *
基于上下文信息融合与注意力感知的目标检测方法研究;谢杰;《中国优秀硕士学位全文数据库信息科技辑》;20200115(第01期);I138-2014 *
基于车辆通行大数据的高速公路路况预测的研究与应用;张振;《中国优秀硕士学位论文全文数据库工程科技II辑》;20210115(第01期);C034-854 *

Also Published As

Publication number Publication date
CN113066544A (en) 2021-07-02

Similar Documents

Publication Publication Date Title
Truong et al. Selective review of offline change point detection methods
Liu et al. Classification of heart diseases based on ECG signals using long short-term memory
Übeyli Combined neural network model employing wavelet coefficients for EEG signals classification
Nejad et al. A new enhanced learning approach to automatic image classification based on Salp Swarm Algorithm
CN116861354A (en) Abnormality identification and mode distinguishing method and system for monitoring data of power transformer
CN111000553A (en) Intelligent classification method for electrocardiogram data based on voting ensemble learning
Güler et al. A modified mixture of experts network structure for ECG beats classification with diverse features
Atal et al. A hybrid feature extraction and machine learning approaches for epileptic seizure detection
Banerjee et al. A hybrid CNN-LSTM architecture for detection of coronary artery disease from ECG
CN113066544B (en) FVEP characteristic point detection method based on CAA-Net and LightGBM
CN107045624B (en) Electroencephalogram signal preprocessing and classifying method based on maximum weighted cluster
CN114004258B (en) Semi-supervised electrocardiographic abnormality detection method
US6941288B2 (en) Online learning method in a decision system
CN111816298A (en) Event prediction method and device, storage medium, terminal and cloud service system
Usha et al. Feature Selection Techniques in Learning Algorithms to Predict Truthful Data
CN114224354B (en) Arrhythmia classification method, arrhythmia classification device, and readable storage medium
Muthulakshmi et al. Prediction of Heart Disease using Ensemble Learning
CN112465054B (en) FCN-based multivariate time series data classification method
Monfared et al. Assessing classical and evolutionary preprocessing approaches for breast cancer diagnosis
CN108615024A (en) A kind of EEG signal disaggregated model based on genetic algorithm and random forest
CN116028858A (en) Pre-labeled self-supervision neural network learning heart beat classification method and system
CN115500790A (en) Method for predicting and analyzing epileptic seizure based on BI-LSTM network
Nejad et al. A novel image categorization strategy based on salp swarm algorithm to enhance efficiency of MRI images
Baek et al. Multi-choice wavelet thresholding based binary classification method
Zhang Machine Learning based Heart Disease Prediction Task

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant